fidoptLiver 3D US-CT
3D ultrasound of the liver (fixed) and contrast-enhanced abdominal CT (moving) of one patient, from two hospitals and two ultrasound devices (the training data of 3D SA-LoIFM).
Read before comparing (8)
- Alignment. prepare.py puts the ultrasound in the CT's scanner frame, by its header only, through the provider's rigid transform from its automatic vessel-based registration. That transform uses no landmark, so the landmarks score it. On case34 it misses every landmark pair by 11 to 35 mm, and the provider leaves case34 out of its own training list: there the least-squares rigid fit on its landmarks replaces it, leaving 5.0 mm.
- Baseline. With no registration the 10 test pairs score a landmark error of 6.2 mm (2.7 to 9.8), a label Dice of 0.35 (0.23 to 0.49) and an HD95 of 39 mm. The provider's frame leaves 1.8 to 10 mm a case, up to 23 mm at a point; on case03 the frame itself is off, by 18 to 21 mm at two points.
- Landmarks. Points on the liver vessels, 2 to 7 a sweep, which the archive calls ground truth without saying who placed them; they look placed by hand in MITK. In 29 pairs the ultrasound point sits where the provider's alignment carries the CT point: those would score that transform only and are left out. Points placed on aligned images lean towards that alignment, which may favour methods that stay near it.
- Few points. A test case keeps 1 to 7 pairs, 38 in all, and three cases rest on one or two: a case's error rests on few points.
- Masks. The provider's, its maker not stated: 1 the liver without its vessels, 2 the gallbladder, 3 the liver vessels. On the ultrasound they cover only the part of the liver the sweep reaches, with straight cuts and islands on some cases, so the CT's liver beyond the sweep counts against every field alike. The overlap is shown, never ranked.
- Grid. The fixed image is the archive's original sweep, with 0.24 to 0.65 mm voxels, and the field is asked on its grid. The labels and landmarks, given only on a copy turned by an average probe pose, are brought back onto it: the landmarks exactly, the labels by nearest neighbour. The fixed mask is where the sweep recorded something.
- Patients. Sweeps that share a CT are one patient, and case13's CT is a second series of case11's patient (their liver masks overlap at Dice 0.99). Which hospital holds which device is not stated. The CT is in Hounsfield units.
- Outside numbers. The 2023 paper reports a mean target registration error of 4.9 mm over these 42 pairs after its rigid registration and 3.0 mm after its B-spline stage; its rigid transforms leave 5.2 mm at the archive's 175 landmark pairs, the left-out ones included. The archive's B-spline fields are not used.
No method has a result on every pair of this dataset yet.
| Accuracy | Smoothness | Cost | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | TRE mm | Change | Dice * | HD95 mm * | rTRE | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| No registration | 6.18 ±1.51 | 0.345 ±0.050 | 38.70 ±11.91 | 0.0207 ±0.0056 | 0.000 | 0.00 | ||||||
* shown, not ranked: on this dataset it is checked against the output of a program.
Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. TRE: distance between matching landmarks. rTRE: the TRE over the length of the image's diagonal, to compare datasets of other sizes and units. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 10 test patients drawn with seed 2026 among the 14 outside development, five from each ultrasound device, and 10 development patients. The archive's 42 sweeps come from 24 patients; each patient keeps one sweep, the one with the most landmark pairs.
- Truth
- landmarks as released contours as released the dataset's masks, 1 the liver (its vessels apart), 2 the gallbladder, 3 the liver vessels, on the ultrasound only where the sweep reaches the dataset's landmark pairs on the liver vessels, those the provider's rigid transform holds left out: 1 to 7 a case
- Source
- huggingface.co/datasets/hbcfx/Liver-3DUS-CT-registration
github.com/hbcfx/3D-SA-LoIFM
doi.org/10.1016/j.bspc.2026.109664
doi.org/10.1002/mp.16396 - Licence
- none stated (no dataset card or licence tag on Hugging Face, none in the archive's readme or the code repository)
10 pairs for tuning, never tested
case41, case39, case09, case37, case07, case35, case15, case31, case19, case32